Release #126

• Revert shallow idle state commit (back to deep sleep)
• Process forked child process BEFORE parent process (lower latency on fork)
• Disable WALT task boosting
• Lock dirty background ratio to exactly half of the dirty ratio
• Process read requests on their origin CPU (if idle) for upwards of threefold performance gain in iozone reader and re-reader tests
• Remove arter's old urandom commit
• Route random fops to urandom (refactored version of arter's; no warnings and no poll)
• Fix compiler warnings
• Use a bitmask for rq_affinity (COMP | FORCE | READ)
• Compile DTB separate from Image.gz
• Fix DTS changes not taking effect
• Adreno 640 napping
• CMA size increase
• Do not expedite RCU
• Remove AVB checks
• Disable TESTMODE for netlink
• Internal rebase and cleanup

Release #105

• 250 Hz timer (4 ms)
• Batch multiple times in Anxiety I/O scheduler
• Use exponential frequency selection in schedutil
• Smooth WALT predicted load
• Use shallow idle state (in testing)
• Normal RCU grace periods after boot
• Heavy rebase
• 7T / 7T Pro support added

Beta #80

• Add custom rom support (use the corresponding build)
• Do extended merges (was off before; turns out RedHat says its still viable for flash)
• Disable autogrouping since Android does a better job
• Patch potential future issue with Anxiety I/O scheduler
• Use 300hz timer for lower latencies (3.33... ms jiffy)
• EXPERIMENTAL: Align scheduling periods with our kernel timer

Without Experimental Commit

• Every 3.33ms, the scheduler gets an interrupt
• If 10ms had passed from the previous cycle, check if we should load balance
  Result: ~6.67ms of times where the scheduler could have moved tasks around, but didn't

With Experimental Commit

• Every 3.33ms, the scheduler gets an interrupt
• Perform a load balance check
  Result: Scheduler can take advantage of new loadavg information and reduce load balancing latency three-fold

Beta #70

• Garbage collect wakelocks
• Reduce number of wakelocks to lock before scanning for idle ones
• Start killing wakelocks right away if the current CPU is idle
• Update Anxiety I/O scheduler
  -- Reduce looping
  -- Remove redundant checks
• Remove high priority flag for thermal ADC
• Use LMKd instead of LMK
• Reduce I/O latency by limiting max requests

Beta #59

• Remove hispeed frequency boosting in schedutil
• Disable RCU expedition
• Remove high priority flags from thermal and dwc3 drivers
• Disable readahead (proven 70% improvement in perf / power)
• Update WiFi to latest CAF
• Halve SCHED_RR (round robin) timeslices

Beta #47

• Disable msm_performance and core_ctl
• Activate schedutil's hispeed frequency at 80% usage instead of 90% (lower direct input latency)
• Process I/O on the origin CPU to score a cache hit
• Set readahead to 128KiB instead of 512KiB (flash memory suffers higher latency with higher readahead)
• Check dirty writebacks every 1MiB rather than every 128 KiB (less jitter and keep cache valid longer)
• Reduce cache pressure to 50% (our RAM size can handle it)
• Load balance less tasks per scheduling cycle (lower latency)
• Compile DTBO in the kernel
• Remove iowait boosting from schedutil

CUSTOM ROMS ARE FIXED
Huge thanks to Podagee and also @Kaz205 for helping me test the builds.

Beta #30

• Removed some commits that could be problematic
• Preserve magisk on install
• Pair a companion module to disable OnePlus brain service on Stock ROM. This fixes the CPU frequency issues.

Why?
The max frequency issue is a result of disabling the Houston kernel driver that is part of OnePlus. I had the option of either sucking it up and keeping the OnePlus bloat, or I could kill the service in userspace. I opted for the latter.

The kernel will now add a magisk module to the device. To use this module, Magisk must be installed prior to the kernel installation.

Beta #23

• Remove AVB checks
• Disable SECCOMP (faster syscalls)
• Disable Audit (expensive in Android)
• Use non-blocking random pool
• Target small cluster in Makefile
• Disable usb_ban and rmnet logging
• Disable IRQ debugging
• Don't use high-priority workqueue for usb policy_engine
• Disable profiling (kdrag0n)
• Anxiety I/O scheduler (sync read priority)
• Schedule last-woken task for better cache locality
• Reduce cache pressure to keep cache valid for longer
• Increase vmstat interval to reduce intermittent jitter
• Adjust dirty ratios to writeback dirty pages later rather than sooner
• Do not collect I/O stats from block subsystem
• Don't dump oom_killed tasks
• Do not use UFS for randomness
• Disable ASLR fully by default (performance not benchmarked)
• Do not attempt extended merge requests since they are CPU heavy
• Make input boosting consistent instead of flickering
• Disable ESD check for panels
• Disable JTAG debugging by default
• Inform kcompact of fragmentation moreso than full RAM usage
• Do not compile HIGHMEM (unused by our system)
• Awaken kswapd at 1% (80mb free left) instead of 0.1% (8mb free left)
• Do not compile watchdog (reduces jitter)
• BFP JIT (better networking performance)
• TCP Fastopen (lower net latency) (kdrag0n)
• Disable binder logging
• Do not print UFS debug information
• Force block requests to execute on their origin cpu (better cache locality)
• Increase haptic strength
• Allow GPU to sleep / nap
• Allow TCP connections to be reused
• Do not compile zRAM (unnecessary for this kernels MM settings)
• Do not compile OnePlus-specific features (memplus, smartboost, etc.) (interferes with kernel)
• Do not perform readahead unless we exceed a read of 64 KiB (do not waste time on trivial reads)
• Configure CFS to allow equal thread timeslices for ten threads at once (in testing)
• Overcommit memory for faster Android start times (allow LMKd to clean up the mess)
• Remove a bunch of debug and tracing that is unnecessary for the kernel
• More that I forgot